net: add skb_defer_max sysctl

commit 68822bdf76 ("net: generalize skb freeing
deferral to per-cpu lists") added another per-cpu
cache of skbs. It was expected to be small,
and an IPI was forced whenever the list reached 128
skbs.

We might need to be able to control more precisely
queue capacity and added latency.

An IPI is generated whenever queue reaches half capacity.

Default value of the new limit is 64.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Documentation/admin-guide/sysctl/net.rst

@@ -322,6 +322,14 @@ a leaked reference faster. A larger value may be useful to prevent false
 warnings on slow/loaded systems.
 Default value is 10, minimum 1, maximum 3600.
 
+skb_defer_max
+-------------
+
+Max size (in skbs) of the per-cpu list of skbs being freed
+by the cpu which allocated them. Used by TCP stack so far.
+
+Default: 64
+
 optmem_max
 ----------
 

net/core/dev.c

@@ -4330,6 +4330,7 @@ int netdev_max_backlog __read_mostly = 1000;
 EXPORT_SYMBOL(netdev_max_backlog);
 
 int netdev_tstamp_prequeue __read_mostly = 1;
+unsigned int sysctl_skb_defer_max __read_mostly = 64;
 int netdev_budget __read_mostly = 300;
 /* Must be at least 2 jiffes to guarantee 1 jiffy timeout */
 unsigned int __read_mostly netdev_budget_usecs = 2 * USEC_PER_SEC / HZ;

net/core/dev.h

@@ -39,7 +39,7 @@ void dev_addr_check(struct net_device *dev);
 /* sysctls not referred to from outside net/core/ */
 extern int		netdev_budget;
 extern unsigned int	netdev_budget_usecs;
-
+extern unsigned int	sysctl_skb_defer_max;
 extern int		netdev_tstamp_prequeue;
 extern int		netdev_unregister_timeout_secs;
 extern int		weight_p;

net/core/skbuff.c

@@ -80,6 +80,7 @@
 #include <linux/user_namespace.h>
 #include <linux/indirect_call_wrapper.h>
 
+#include "dev.h"
 #include "sock_destructor.h"
 
 struct kmem_cache *skbuff_head_cache __ro_after_init;
@@ -6496,16 +6497,21 @@ void skb_attempt_defer_free(struct sk_buff *skb)
 	int cpu = skb->alloc_cpu;
 	struct softnet_data *sd;
 	unsigned long flags;
+	unsigned int defer_max;
 	bool kick;
 
 	if (WARN_ON_ONCE(cpu >= nr_cpu_ids) ||
 	    !cpu_online(cpu) ||
 	    cpu == raw_smp_processor_id()) {
-		__kfree_skb(skb);
+nodefer:	__kfree_skb(skb);
 		return;
 	}
 
 	sd = &per_cpu(softnet_data, cpu);
+	defer_max = READ_ONCE(sysctl_skb_defer_max);
+	if (READ_ONCE(sd->defer_count) >= defer_max)
+		goto nodefer;
+
 	/* We do not send an IPI or any signal.
 	 * Remote cpu will eventually call skb_defer_free_flush()
 	 */
@@ -6515,11 +6521,8 @@ void skb_attempt_defer_free(struct sk_buff *skb)
 	WRITE_ONCE(sd->defer_list, skb);
 	sd->defer_count++;
 
-	/* kick every time queue length reaches 128.
+	/* Send an IPI every time queue reaches half capacity. */
-	 * This condition should hardly be hit under normal conditions,
+	kick = sd->defer_count == (defer_max >> 1);
-	 * unless cpu suddenly stopped to receive NIC interrupts.
-	 */
-	kick = sd->defer_count == 128;
 
 	spin_unlock_irqrestore(&sd->defer_lock, flags);
 

net/core/sysctl_net_core.c

@@ -578,6 +578,14 @@ static struct ctl_table net_core_table[] = {
 		.extra1		= SYSCTL_ONE,
 		.extra2		= &int_3600,
 	},
+	{
+		.procname	= "skb_defer_max",
+		.data		= &sysctl_skb_defer_max,
+		.maxlen		= sizeof(unsigned int),
+		.mode		= 0644,
+		.proc_handler	= proc_dointvec_minmax,
+		.extra1		= SYSCTL_ZERO,
+	},
 	{ }
 };